Today: Dummy Agent Library — the first hands-on one. Ran it in Colab against Hugging Face's serverless InferenceClient, no agent framework at all, just raw chat completions.
Building the loop manually made Days 3–6 concrete. You write a system prompt describing the tools and the required output format, send it, get back an action, parse it, run the function yourself, append the result, send the whole thing again.
The detail that actually surprised me: without a stop sequence, the model happily writes the Observation: itself — inventing a tool result that never happened and confidently continuing from it. The fix is stop=["Observation:"], which halts generation right at the boundary so you can execute the real tool and inject the real output before resuming.
That's the "stop and parse" idea from Day 5, except now I've seen the failure mode it prevents. The model has no way to know whether a tool actually ran — it's all just text continuation, so left alone it will cheerfully autocomplete a plausible-looking result.
Also fair to say the exercise is deliberately tedious. Hand-managing message lists and string parsing for a single weather call is exactly the argument for smolagents, which is up next.